检索结果-内蒙古大学图书馆

UformPose: A U-Shaped Hierarchical Multi-Scale keypoint-Aware Framework for Human Pose Estimation

IEEE Transactions on Circuits and Systems for Video Technology 2023年第4期33卷 1697-1709页

作者： Wang, You-Jie Luo, Yan-Min Bai, Gui-Hu Guo, Jing-Ming Huaqiao University Key Laboratory for Computer Vision and Pattern Recognition The College of Computer Science and Technology Xiamen361021 China National Taiwan University of Science and Technology Department of Electrical Engineering Taipei10607 Taiwan

Human pose estimation is a fundamental yet challenging task in computer vision. However, difficult scenarios such as invisible keypoints, occlusions and small-scale persons are still not well-handed. In this paper, we present a novel pose estimation framework named UformPose which targets to relieve these issues. UformPose has two core designs: Shared Feature Pyramid Stem (SFPS) and U-shaped hierarchical Multi-scale keypoint-aware Attention Module (U-MKAM). SFPS is a feature pyramid stem with shared mechanism to learn stronger low-level features at the initial stage, and the shared mechanism can facilitate cross-resolution commonality learning. Our U-MKAM attempts to generate high-quality high-resolution representations by integrating all levels of feature representation of the backbone layer by layer. More importantly, we utilize the flexibility of attention operations for keypoint-aware modeling, which explicitly captures and trades-offs the dependencies between keypoints. We empirically demonstrate the effectiveness of our framework through the competitive pose estimation results on the COCO dataset. Extensive experiments and visual analysis on CrowdPose demonstrate the robustness of our model in crowd scenes. © 1991-2012 IEEE.

关键词： computer vision

来源：评论

学校读者我要写书评

暂无评论

Revisiting the Generalization Problem of Low-level vision Models Through the Lens of Image Deraining

arXiv

引用

arXiv 2025年

作者： Hu, Jinfan You, Zhiyuan Gu, Jinjin Zhu, Kaiwen Xue, Tianfan Dong, Chao Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences Shenzhen518055 China University of Chinese Academy of Sciences Beijing100049 China The Chinese University of Hong Kong 999077 Hong Kong The University of Sydney NSW2006 Australia Shanghai Jiao Tong University Shanghai200240 China Shanghai Artificial Intelligence Laboratory Shanghai200232 China Shenzhen Key Lab of Computer Vision and Pattern Recognition Shenzhen Institutes of Advanced Technology Chinese Academy of Sciences China Shenzhen University of Advanced Technology Shenzhen518055 China

Generalization remains a significant challenge for low-level vision models, which often struggle with unseen degradations in real-world scenarios despite their success in controlled benchmarks. In this paper, we revisit the generalization problem in low-level vision models. Image deraining is selected as a case study due to its well-defined and easily decoupled structure, allowing for more effective observation and analysis. Through comprehensive experiments, we reveal that the generalization issue is not primarily due to limited network capacity but rather the failure of existing training strategies, which lead networks to overfit specific degradation patterns. Our findings show that guiding networks to focus on learning the underlying image content, rather than the degradation patterns, is key to improving generalization. We demonstrate that balancing the complexity of background images and degradations in the training data helps networks better fit the image distribution. Furthermore, incorporating content priors from pre-trained generative models significantly enhances generalization. Experiments on both image deraining and image denoising validate the proposed strategies. We believe the insights and solutions will inspire further research and improve the generalization of low-level vision models. Copyright © 2025, The Authors. All rights reserved.

关键词： Image denoising

来源：评论

学校读者我要写书评

暂无评论

GCAT: graph calibration attention transformer for robust object tracking

引用

Neural Computing and Applications 2024年第23期36卷 14151-14172页

作者： Chen, Si Hu, Xinxin Wang, Da-Han Yan, Yan Zhu, Shunzhi Fujian Key Laboratory of Pattern Recognition and Image Understanding School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China State Key Laboratory of Integrated Services Networks Xidian University Xi’an710071 China Fujian Key Laboratory of Sensing and Computing for Smart City School of Informatics Xiamen University Xiamen361005 China

Recent Siamese trackers have taken advantage of transformers to achieve impressive advancements. However, existing transformer trackers ignore considering the positional and structural information between tokens, and traditional template update strategies easily introduce noises to the dynamic templates during tracking. In order to alleviate this issue, this paper develops a novel end-to-end graph calibration attention transformer network (GCAT) to enhance tracking robustness and accuracy. A graph calibration attention mechanism is first designed to calibrate and aggregate template information, for effectively updating dynamic templates during the tracking process. Specifically, each token is considered as a node in the graph, and then, we calculate the weight relationships between each node and their adjacent nodes. Thus, this mechanism can aggregate the global context information of the template and search nodes and activate feature channels based on weights and biases to obtain more discriminative feature information. Moreover, we leverage a multi-level dropout mechanism to perform the data dropout, the layer dropout, and the feature dropout on the data, network, and attention levels, respectively, to avoid overfitting of local-specific information and improve the generalization ability. Extensive experiments show the proposed method achieves superior performance on seven challenging benchmark datasets, i.e., OTB100, OTB2013, UAV123, LaSOT, GOT10K, VOT2020, and TrackingNet. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024.

关键词： Aggregates

来源：评论

学校读者我要写书评

暂无评论

Deform-CAM: Self-attention Based on Deformable Convolution for Weakly Supervised Semantic Segmentation 17th

Deform-CAM: Self-attention Based on Deformable Convolution f...

引用

17th Chinese Conference on Image and Graphics Technologies and Applications, IGTA 2022

作者： Huang, Feihong Wang, Da-Han Ye, Hai-Li Zhu, Shunzhi School of Computer and Information Engineering Xiamen University of Technology Xiamen China Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen China

ISBN: (纸本)9789811950957

Weakly-supervised semantic segmentation (WSSS) receives increasing attentions from the community in recent years as it leverages the weakly annotated data to solve the problem of lacking of fully annotated data. Among them, the WSSS method based on image-level annotation is the most direct and effective while the image-level annotation is easy to obtain. Most advanced methods use class activation maps (CAM) as initial pseudo-labels, however, they only identify local regions of the target, while ignoring the context information among local regions. To solve this problem, this paper proposes a deformable convolution based self-attention module (DSAM), which introduces a pixel relationship matrix, to learn the contextual information of the image. A regularization loss is introduced to narrow the distance between the DSAM and the CAM. Compared to the base CAM method, our method can identify more target features and robustly improve the performance of WSSS without training the classifier multiple times. Our proposed method achieves the mIoU of 65.5% and 66.8% on the Pascal VOC 2012 val and test sets, respectively, demonstrating the feasibility of the method. © 2022, The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd.

关键词： Convolution

来源：评论

学校读者我要写书评

暂无评论

Multi-Zone Transformer Based on Self-Distillation for Facial Attribute recognition 17

Multi-Zone Transformer Based on Self-Distillation for Facial...

引用

17th IEEE International Conference on Automatic Face and Gesture recognition, FG 2023

作者： Chen, Si Zhu, Xueyan Wang, Da-Han Zhu, Shunzhi Wu, Yun Fujian Key Laboratory of Pattern Recognition and Image Understanding School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China

ISBN: (纸本)9798350345445

Recently, transformers have shown great promising performance in various computer vision tasks. However, the current transformer based methods ignore the information exchanges between transformer blocks, and they have not been applied in the facial attribute recognition task. In this paper, we propose a multi-zone transformer based on self-distillation for FAR, termed MZTS, to predict the facial attributes. A multi-zone transformer encoder is firstly presented to achieve the interactions of the different transformer encoder blocks, thus avoiding forgetting the effective information between the transformer encoder block groups during the iteration process. Furthermore, we introduce a new self-distillation mechanism based on class tokens, which distills the class tokens obtained from the last transformer encoder block group to the other shallow groups by interacting with the significant information between the different transformer blocks through attention. Extensive experiments on the challenging CelebA and LFWA datasets have demonstrated the excellent performance of the proposed method for FAR. © 2023 IEEE.

关键词： Face recognition

来源：评论

学校读者我要写书评

暂无评论

Exploring Joint Information of Multi-Scales for Vehicle Re-Identification 22

Exploring Joint Information of Multi-Scales for Vehicle Re-I...

引用

8th International Conference on Computing and Artificial Intelligence, ICCAI 2022

作者： Zhou, Yongjie Wang, Dahan School of Computer and Information Engineering Xiamen University of Technology Xiamen Chinese Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China

ISBN: (纸本)9781450396110

Vehicle re-identification (re-ID) is an essential component of intelligent video surveillance, which attempts to solve the problem of retrieving specific vehicle instances. The technical challenge is mainly the requirement that the algorithm be robust under different viewing angles, resolutions, occlusions, and lighting conditions, yet different conditions can arise where the appearance of the exact vehicle varies dramatically. In contrast, the appearance of two different vehicles is exceptionally similar. To overcome this enormous challenge: we adopt a novel multi-scales attention network architecture, which will feature representation for feature maps at different scales to enhance the recognition of intricate parts, and introduce an inter-coordinate attention module to build interdependencies between component positions, making it possible to build a more robust vehicle model feature representation, and finally this paper also proposes a camera-based vehicle re localization method that enables the model to calculate the vehicle occurrence probability using the potential information among road network cameras as a post-processing step to re-correct the confidence level of the re-identification ranking. Experimental results show that the method outperforms the reference method in all indexes even without post-processing. At the same time, the camera confidence correction can be used in overlay with the reordering clustering correction method, which can improve the results by four percentage points and MAP up to 81.4%. © 2022 ACM.

关键词： Vehicles

来源：评论

学校读者我要写书评

暂无评论

Multi-task learning with deep dual-path network for facial attribute recognition 9

Multi-task learning with deep dual-path network for facial a...

引用

9th International Conference on Computing and pattern recognition, ICCPR 2020

作者： Lai, Xinyu Chen, Si Wang, Da-Han Zhu, Shunzhi School of Computer and Information Engineering Xiamen University of Technology Fujian Key Laboratory of Pattern Recognition and Image Understanding China

ISBN: (纸本)9781450387835

Facial attribute recognition is a popular and challenging research topic in computer vision. In the traditional deep learning based attribute recognition methods, the mid-level network features and the differences between attribute groups are not fully explored. To solve the above problem, a deep dual-path network is proposed for facial attribute recognition. In the multi-task learning framework, two sub-networks are employed to respectively extract the features of two attribute groups, i.e., local attributes and global ones, and designed with both different scale images and different depth networks. Furthermore, an adaptive Focal loss penalty scheme is developed to automatically assign weights to handle the class imbalance problem for facial attribute recognition. Experimental results on the challenging CelebA dataset show that the proposed method achieves the better performance than state-of-the-art methods. © 2020 ACM.

关键词： Deep learning

来源：评论

学校读者我要写书评

暂无评论

Handwritten mathematical expression recognition with self-attention 4

Handwritten mathematical expression recognition with self-at...

引用

4th International Conference on Algorithms, Computing and Artificial Intelligence, ACAI 2021

作者： Chi, Xueke Wang, Da-Han Wu, Yuefeng Wu, Yun Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China

ISBN: (纸本)9781450385053

Attention-based encoder-decoder models have made great success on handwritten mathematical expression recognition in recent years. However, this kind of method has the problem of attention drift, because under the local attention mechanism based on RNN, the high similarity between coding features can cause attention confusion. To settle this problem, we propose an encoder-decoder model with self-attention, which captures the global information of the feature map and fuses the local information of the CNN as complementary features. Experiments are conducted on the CROHME2014 and CROHME 2016 competition datasets. The experimental results show that, when only using the official training dataset, the proposed method achieves recognition accuracies of 51.98% and 50.74% on the CROHME2014 and CROHME2016 competition datasets, respectively, which outperforms the other methods significantly. The improvements demonstrate the effectiveness of the self-attention module. © 2021 ACM.

关键词： Signal encoding

来源：评论

学校读者我要写书评

暂无评论

Dynamic Updating of Partial Feature Maps for Image-based Real-Time recognition of Online Handwritten Chinese Text 22

Dynamic Updating of Partial Feature Maps for Image-based Rea...

引用

8th International Conference on Computing and Artificial Intelligence, ICCAI 2022

作者： Wu, Yuefeng Zhong, Ying Wang, Da-Han Chi, Xueke School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China

ISBN: (纸本)9781450396110

Online handwritten Chinese text recognition (OHCTR) has made certain progress in recent years. Most methods currently used are image-based text recognition which has achieved some success. However, it should be noticed that, it is difficult to smoothly recognize online handwritten texts in real-time, especially for long text lines. For real-time recognition of OHCT, the feature extraction of ever-extending text lines will be time-consuming and make the system un-humanization. Moreover, the feature extraction of the non-written part will be performed repeatedly, which is not suitable and has not been addressed by previous works. In this paper, we propose a method to dynamically update partial feature maps of the newly written part for real-time OHCTR. The basic idea is to update the feature maps according to the newly written stroke, which can avoid repeated convolutional computations and thus improve the recognition speed. The proposed method can handle the case of delayed strokes, and can greatly improve the speed of real-time recognition without any loss of accuracy. Experiments on the CASIA-OLHWDB and the ICDAR-2013 competition database demonstrate that our real-time recognition speed are multiplier increased compared with traditional text line recognition methods. © 2022 Association for Computing Machinery. All rights reserved.

关键词： Convolution

来源：评论

学校读者我要写书评

暂无评论

CDE-Net: A semi-supervised nucleus instance segmentation method based on center deviation estimation 22

CDE-Net: A semi-supervised nucleus instance segmentation met...

引用

8th International Conference on Computing and Artificial Intelligence, ICCAI 2022

作者： Yan, Yi Wang, Da-Han Ye, Haili Zhu, Shunzhi Li, Jianmin School of Computer and Information Engineering Xiamen University of Technology Xiamen361024 China Fujian Key Laboratory of Pattern Recognition and Image Understanding Xiamen361024 China

ISBN: (纸本)9781450396110

Cell nucleus segmentation is an important step in the diagnosis of tumor malignancy. In the case of out-of-focus, the cell nucleus boundary tends to be blurred, which increases the difficulty of cell segmentation and marking. Therefore, in the case of marking only some core instances, it is very important to realize the segmentation of core instances. In this paper, we propose a semi-supervised nucleus instance segmentation method based on anchor bias prediction that is achieved by our proposed center deviation estimation network (CDE-Net). By introducing the CDE branch to predict the offset degree of the anchor frame, and weighting the confidence of the anchor frame, more accurate prediction results can be obtained. The model adopts the semi-supervised training method to progressively learn the CDE-Net model, which alleviates the problem of imbalance between positive and negative samples caused by only marking part of the nuclear samples. The experimental results show that CDE-Net is superior to Mask R-CNN in many evaluation indicators, and achieves SOTA on the partially labelled MoNuSeg data set. © 2022 Association for Computing Machinery. All rights reserved.

关键词： Diagnosis

来源：评论

学校读者我要写书评

暂无评论

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案：

请选择收藏分类：

通借通还

建议与咨询 留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

时间限定

文献类型

馆藏选择

核心期刊

语言

文献类型

帮助

文字说明：

检索规则说明：

检索范例：

分类表

所选分类

限定检索结果

文献类型

馆藏范围

日期分布

学科分类号

主题

机构

作者

语言

请选择保存的检索档案： 新增检索档案 确定 取消

请选择收藏分类： 新增自定义分类 确定 取消

通借通还

建议与咨询留下您的常用邮箱和电话号码，以便我们向您反馈解决方案和替代方法

请选择保存的检索档案：

请选择收藏分类：